Universal Image Vaccine Against Steganography

Abstract

In the past decade, the diversification of steganographic techniques has posed significant threats to information security, necessitating effective countermeasures. Current defenses, mainly reliant on steganalysis, struggle with detection accuracy. While “image vaccines” have been proposed, they often target specific methodologies. This paper introduces a universal steganographic vaccine to enhance steganalysis accuracy. Our symmetric approach integrates with existing methods to protect images before online dissemination using the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. Experimental results show significant accuracy improvements across traditional and deep learning-based steganalysis, especially at medium-to-high payloads. Specifically, for payloads of 0.1–0.5 bpp, the original detection error rate was reduced from 0.3429 to 0.2346, achieving an overall average reduction of 31.57% for traditional algorithms, while the detection success rate of deep learning-based algorithms can reach 100%. Overall, integrating CLAHE as a universal vaccine significantly advances steganalysis.

1. Introduction

The widespread adoption of multimedia technology has enhanced daily life but also spurred advancements in digital information hiding and encryption. Concealing secret information during communication remains a critical research focus for governments and enterprises. Prominent information hiding techniques include steganography and watermarking [1,2]. While watermarking is primarily employed for copyright protection and the authentication of digital media, steganography facilitates covert communication by embedding secret information within public digital media such as images, text, audio, and video, without attracting the attention of third parties [3,4,5,6]. For example, Li et al. proposed a steganography algorithm based on VVC, which embeds secret information by modifying the partitioning mode of chroma coding units and uses convolutional neural networks to optimize video quality and bitrate [7]. In contrast to cryptography, which encrypts secret information, steganography not only safeguards the information but also conceals the very existence of the communication [8,9].

In the digital era, especially with cloud computing, digital images uploaded and shared by users have become crucial for disseminating and expressing information [10]. The rapid advancement of technology has made digital images indispensable across various fields, including education, research, national security, and personal life. For instance, the HPDH-MI algorithm enhances data transmission security in the e-health domain by embedding secret data in AMBTC-compressed medical images, achieving high payload capacity, image quality, and embedding efficiency, making it ideal for telemedicine applications [11]. Tuo et al. integrated audio encryption with information hiding in image carriers, utilizing multidimensional chaotic maps and GHM multi-wavelet transforms to achieve robust security and diverse applications [12]. Consequently, they have become the primary medium for steganography [13].

Modern steganography algorithms commonly embed secret information into areas of an image that are complex in texture and difficult to detect. This approach minimizes distortion, making changes visually indistinguishable and significantly enhancing the concealment of information [14,15,16,17,18]. These methods are primarily based on content-adaptive steganography designed within the framework of minimizing distortion [19]. By calculating a distortion cost function, they identify pixel locations with the lowest embedding cost and use encoding techniques like Syndrome-Trellis Codes (STCs) [20] to minimize the distortion function, resulting in the final stego image. With the rise in neural networks, deep learning-based steganography algorithms have emerged, such as UT-GAN [21], which employs generative adversarial networks to improve the stealth of information embedding while maintaining high image quality and robustness against steganalysis.

However, the misuse of steganography poses a serious threat to information security and personal privacy. For instance, malicious organizations distribute altered digital images and other multimedia files over the internet to conduct illegal activities [22]. Therefore, researching effective steganography defense strategies is crucial for protecting the information security of digital images transmitted over public channels, leading to the development of steganalysis techniques. Steganalysis primarily relies on analyzing statistical feature differences in images to identify steganographic activities. Traditional steganalysis methods [23,24,25,26,27] capture subtle statistical feature changes in images by constructing rich models and using ensemble classifiers to detect steganography. Since Xu et al. designed a convolutional neural network structure to enhance detection capabilities in steganalysis [28], deep learning-based steganalysis methods have been continuously emerging [29,30]. However, as steganography advances, the accuracy of steganography detection still requires improvement.

In the latest work, the SUDS framework utilizes deep learning techniques to detect and eliminate steganographic content, focusing specifically on the removal of hidden information [31]. In contrast, Chen et al.’s “Universal Watermark Vaccine” employs universal adversarial perturbations for watermark protection, prioritizing watermark security over steganography defense [32]. The “image vaccine” steganography defense mechanism, initially proposed by Li et al., involves injecting vaccine data by modifying pixels and Discrete Cosine Transform (DCT) coefficients [33]. However, its effectiveness is restricted to specific steganography methods. The limitations mentioned have prompted us to develop a more universal solution to safeguard images against the threat of various steganographic techniques. In response, we introduce an innovative “universal image vaccine” to boost anti-steganography capabilities. As shown in Figure 1, injecting this vaccine into images allows the system to more easily detect hidden information.

In practical applications, the majority of digital images are generated from photos taken by users with mobile phones or cameras. Thus, applying the image vaccine as a filter during image capture not only enhances image quality but also bolsters the anti-steganography capabilities of digital images. Our method seamlessly integrates with all steganalysis techniques to protect images prior to their upload to public channels, thereby making steganographic activities more detectable and identifiable by existing detection technologies. This approach effectively enhances the overall detection capability of steganography detection systems.

Our approach does not modify existing steganalysis techniques. Instead, it fortifies images before users upload them to public channels, facilitating the detection of steganographic activities by existing methods and significantly improving the efficacy of steganalysis.

The main contributions of this paper are as follows:

• We propose a “universal image vaccine” method against steganography. This approach optimizes image quality while improving the accuracy of steganalysis to varying degrees, effectively detecting steganographic operations. When combined with machine learning algorithms, it can achieve a detection success rate of up to 100%.
• Our method utilizes the CLAHE [34] algorithm as an image vaccine, applicable to all types of steganography and steganalysis methods, including both traditional and deep learning approaches, offering a wide range of applications.

The subsequent sections of this paper are structured as follows: Section 2, ‘Method’, introduces our image vaccine scheme; Section 3, ‘Experimental Results’, presents the experimental setup as well as the results and analysis; and Section 4, ‘Conclusions’, summarizes the research findings.

2. Method

Before uploading photos to cloud environments, users often employ optimization algorithms to enhance image quality. As illustrated in Figure 2, we address practical needs by selecting the CLAHE algorithm as a “universal image vaccine” technique. This is integrated with existing steganalysis methods to create a comprehensive defense system, thereby improving overall steganalysis performance. Specifically, we apply the CLAHE algorithm to original digital images, which not only enhances visual quality but also results in protected digital images.

2.1. The Influence of Image Vaccines on Steganalysis

The “image vaccine” concept, initially proposed by Li et al., [33] presents certain limitations when applied to real-world scenarios. Its efficacy is confined to detection by particular methods, which restricts its broader utility in the field of steganography defense. Recognizing this, we embarked on a quest to develop a universal steganographic vaccine, one that could provide a robust defense against a spectrum of steganographic attacks, including those based on traditional methods as well as sophisticated machine learning algorithms.

Our initial research investigated the effects of image preprocessing on steganalysis. We evaluated basic techniques like sharpening and Wiener filtering on steganographic detectability. While these methods increased detectability, they variably impacted visual quality, rendering them unreliable as a robust countermeasure. We also explored brightness and contrast enhancement, a common image optimization practice, to assess its influence on steganalysis.

We applied traditional histogram equalization (HE) to the 1.pgm and 128.pgm images from the BOWS2 [35] dataset. Referring to Figure 3, this resulted in noticeable color blocks in the sky and significant deviations from the original images. Our experiments indicate that improvements in detection performance often come at the expense of visual quality. Therefore, to enhance detection performance while improving visual quality, we opted for the CLAHE algorithm as an image vaccine method.

2.2. Contrast Limited Adaptive Histogram Equalization

The traditional histogram equalization (HE) algorithm enhances the contrast and clarity of an image, making it appear more vivid and bright. This algorithm works by redistributing the pixel values of the image to achieve a more uniform distribution across the entire range. The process for applying HE to an L-level grayscale image involves the following steps:

• Compute the gray-level histogram: Count the number of pixels $n_k$ for each gray level $r_k$, where k ranges from 0 to $L-1$, and L is the total number of gray levels (typically 256).
• Calculate the probability density function (PDF):

  $$p\left(r_k\right)={\displaystyle \frac{n_k}{N}}$$

  (1)

  where $p\left(r_k\right)$ represents the probability of the k-th gray level, and N is the total number of pixels in the image. The grayscale histogram can be viewed as a probability distribution function.
• Compute and normalize the cumulative distribution function (CDF) and map the gray levels:

  $$s_k=(L-1)\sum _{j=0}^{k}p\left(r_j\right)$$

  (2)
  This step transforms each pixel’s original gray level $r_k$ to the new gray level $s_k$, resulting in the equalized image. Through histogram equalization (HE), the CDF is adjusted to become uniform across its domain, extending the dynamic range of grayscale levels and suppressing those with fewer pixels.

While HE can significantly enhance contrast, it often leads to issues such as noise amplification and inadequate preservation of image details. This occurs because the uniform distribution may excessively enhance contrast in areas with fewer pixel variations, resulting in a loss of detail and increased noise. CLAHE algorithm addresses many of these limitations.

CLAHE is an advanced image processing technique designed to enhance local contrast in images. It works by dividing the image into multiple small blocks and performing histogram equalization independently on each region. This localized approach effectively highlights image details, as each region’s grayscale histogram is computed separately. To mitigate noise amplification, CLAHE employs a contrast-limiting mechanism. This involves capping the pixel count for each grayscale level within a predefined threshold and redistributing excess pixels, which is crucial for preventing excessive contrast enhancement.

Once the histogram is clipped, it is equalized to enhance local contrast within each block. Bilinear interpolation is then used to merge the processed blocks, ensuring smooth transitions between adjacent regions. The enhanced image retains more details, with smoother transitions in areas like the sky, avoiding the unnatural edge transitions common in traditional histogram equalization. This method significantly improves visual quality and is particularly effective in fields such as medical imaging, remote sensing image analysis, and industrial inspection.

2.3. Impact of CLAHE on Visual Quality

To more intuitively assess the impact of CLAHE on image quality, and to comprehensively analyze the differences between HE and CLAHE, we used metrics such as Mean-Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and information entropy for quantitative evaluation. MSE quantifies error, PSNR evaluates overall image quality, SSIM measures perceptual similarity, and information entropy reflects changes in information content. The respective formulas are provided below:

$$\mathrm{MSE}={\displaystyle \frac{1}{mn}}\sum _{i=1}^m\sum _{j=1}^n{[I(i,j)-K(i,j)]}^2$$

(3)

$$\mathrm{PSNR}=10·{log}_{10}\left({\displaystyle \frac{{255}^2}{\mathrm{MSE}}}\right)$$

(4)

where m and n denote the dimensions of the image, and $I(i,j)$ and $K(i,j)$ represent the pixel values of the original and the processed images, respectively. The higher the PSNR value, the less the loss in image quality.

$$\mathrm{SSIM}(x,y)={\displaystyle \frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}}$$

(5)

where ${\mu }_x$ and ${\mu }_y$ represent the mean intensities of images x and y, respectively. The terms ${\sigma }_x^2$ and ${\sigma }_y^2$ denote the variances in these images, and ${\sigma }_{xy}$ signifies the covariance between them. To ensure numerical stability and prevent division by zero or extremely small values, which could result in undefined or infinite SSIM values, constants $C_1$ and $C_2$ are incorporated into the SSIM calculation.

$$H\left(X\right)=-\sum _{i=1}^{n}p\left(x_i\right){log}_{2}p\left(x_i\right)$$

(6)

where $p\left(x_i\right)$ denotes the probability of the grayscale level $x_i$, and n is the total number of grayscale levels. These metrics collectively offer a robust framework for the quantitative analysis of image quality, essential in various applications ranging from compression to enhancement. The SSIM value ranges from 0 to 1, with values closer to 1 indicating better visual quality of the image.

By combining these metrics, we can thoroughly assess the performance differences of the two algorithms in image processing. We calculated and averaged various metrics using 10,000 images from BOSSbase ver. 1.01. In this study, we employed a 16 × 16 grid of local regions and a clip limit of 0.005 for the CLAHE algorithm to optimize image contrast while suppressing noise amplification. As illustrated in

Table 1. Comparison of HE and CLAHE.

	Original	HE	CLAHE
MSE	-	4826.67	502.04
PSNR	-	12.94 dB	21.60 dB
SSIM	-	0.62	0.87
Information Entropy	6.80	5.72	7.17

, in comparison with the HE method, images processed utilizing the CLAHE algorithm demonstrate superior performance metrics, evidenced by higher PSNR and SSIM values, along with reduced MSE values, indicative of a decreased number of pixel alterations. These findings suggest that CLAHE processing is more effective in preserving image details and enhancing image quality, thereby more adeptly fulfilling the requirements of users in practical applications.

2.4. Enhancing Detection Performance

In steganalysis systems, which are framed as binary classification problems, evaluating the performance of detection algorithms is crucial. A key metric in this context is the minimum average decision error rate ($P_E$), defined as

$$P_E={\displaystyle \frac{1}{2}}\underset{P_{FA}}{min}(P_{FA}+P_{MD})$$

(7)

Here, $P_{FA}$ denotes the False Alarm Rate, which refers to the incorrect classification of an image as containing steganography. $P_{MD}$ represents the Missed Detection Rate, indicating the failure to detect actual steganographic content. A lower $P_E$ indicates superior steganalysis detection performance.

The calculation of $P_E$ considers the trade-off between these two fundamental error types in steganalysis, aiming to minimize their sum. This metric is particularly useful for systems that need to balance reducing false positives with improving steganography detection efficiency. By optimizing $P_E$, the accuracy and reliability of steganalysis tools can be significantly enhanced, leading to more effective identification and defense against steganographic attacks.

By applying protective operations to images, we can optimize the $P_E$ value to some extent, thereby resisting steganographic attacks. To further validate the performance of our proposed defense method, we conducted multiple experiments, demonstrating that digital images treated with the “image vaccine” exhibit stronger resistance in steganalysis, regardless of the type of steganographic operation. The specific experimental details will be presented in Section 3.

3. Experimental Results

3.1. Experimental Setup

To validate the effectiveness of our approach, we conducted multiple sets of controlled experiments. We selected various types of images and datasets for training and testing, including LIRMMBase$256\times 256$, BOSSbase ver. 1.01 [36] and BOWS2 [35]. Since the images in BOSSbase and BOWS2 are $512\times 512$ grayscale images, and the UT-GAN [21] steganography algorithm requires input images of $256\times 256$, we resized all original images to $256\times 256$. This ensures the control of variables and facilitates the experiments. In this way, we can more accurately assess the performance of our method across different datasets.

Table 2. Details of the datasets.

Dataset	Volume	Bit Depth	Size	Format
BOSSBase 1.01	10,000	8	$512\times 512$	PGM
BOWS2	10,000	8	$512\times 512$	PGM
LIRMMBase256 × 256	1008	8	$256\times 256$	PGM

summarizes the key characteristics of the datasets used in this study. These datasets are crucial for evaluating the performance of image vaccine processing algorithms, particularly in the context of steganography and steganalysis. Each dataset consists of grayscale images with varying resolutions and formats, which are commonly used benchmarks in the field.

We injected our image vaccine into the original dataset to obtain a protected dataset and conducted experiments to evaluate its effectiveness. Specifically, we selected varying numbers of original images from each dataset for the injection process. To assess the broad applicability of our method to steganography, we used popular steganographic methods WOW [15], S-UNIWARD [16], MiPOD [18], and the machine learning method UT-GAN for data embedding at payloads of 0.1, 0.2, 0.3, 0.4, and 0.5 bpp. The effectiveness of the proposed algorithm in enhancing steganalysis systems was then evaluated by applying existing steganalysis methods for detection and comparing the resulting detection error probabilities ($P_E$, defined in Equation (3)).

The LIRMMBase256 × 256 datasets were used in the training of the UT-GAN steganography algorithm. We trained multiple models for different payloads and used these models to generate the required stego image datasets.

3.2. Analysis of Experimental Results

To evaluate the effectiveness of our proposed method under different steganalysis techniques, we compared traditional and machine learning-based steganalysis methods. We employed traditional feature extraction techniques such as SPAM [23], SRM [24], SRMQ1 [26], and TLBP [27] to extract feature sets from both original and stego images. These feature sets were then evaluated using an ensemble classifier widely used in steganalysis. Additionally, we tested the machine learning-based Yedroudj-Net [30] algorithm.

In our experimental design, half of the original and stego images were used for model training, with the remainder used for testing.

In this experiment, as depicted in Figure 4 on the following page, we first compared our proposed method with traditional steganalysis methods using the BOWS2 dataset. The figure illustrates the performance differences between our method and traditional methods (SPAM, SRM, SRMQ1, TLBP) on 1000 random images from the BOWS2 dataset, across various steganographic techniques and payload rates (ranging from 0.1 bpp to 0.5 bpp). Please refer to Figure 4 for a visual representation of these comparisons. Detailed numerical results are presented in

Table 3. Comparison of our method with traditional steganalytic methods on BOWS2.

		Payload (bpp)
Steganography	Steganalysis	0.1		0.2		0.3		0.4		0.5
		Original	Proposed	Original	Proposed	Original	Proposed	Original	Proposed	Original	Proposed
WOW	SPAM	0.4969	0.4593	0.4755	0.3933	0.4571	0.4244	0.4244	0.2652	0.3888	0.2211
	SRM	0.4784	0.4453	0.4339	0.3674	0.3795	0.2923	0.3336	0.2241	0.2918	0.1743
	SRMQ1	0.4804	0.4394	0.4410	0.3505	0.3970	0.2714	0.3559	0.2080	0.3106	0.1610
	TLBP	0.4829	0.4610	0.4505	0.3928	0.4012	0.3285	0.3599	0.2620	0.3165	0.2062
S-UNIWARD	SPAM	0.4940	0.4467	0.4678	0.3662	0.4367	0.2943	0.4103	0.2311	0.3725	0.1788
	SRM	0.4760	0.4365	0.4284	0.3379	0.3782	0.2577	0.3263	0.1942	0.2813	0.1456
	SRMQ1	0.4802	0.4205	0.4299	0.3190	0.3872	0.2385	0.3388	0.1853	0.2931	0.1361
	TLBP	0.4848	0.4489	0.4472	0.3707	0.4058	0.2908	0.3472	0.2166	0.3032	0.1677
MiPOD	SPAM	0.4937	0.4280	0.4868	0.3461	0.4667	0.2853	0.4444	0.2311	0.4209	0.1873
	SRM	0.4772	0.4059	0.4436	0.3173	0.4097	0.2421	0.3659	0.1882	0.3269	0.1493
	SRMQ1	0.4868	0.3911	0.4518	0.2962	0.4199	0.2258	0.3812	0.1712	0.3428	0.1365
	TLBP	0.4839	0.4187	0.4620	0.3372	0.4248	0.2661	0.3852	0.2103	0.3478	0.1716
UT-GAN	SPAM	0.2598	0.1379	0.1977	0.0746	0.1662	0.0517	0.1576	0.0442	0.1503	0.0356
	SRM	0.1878	0.1250	0.1241	0.0645	0.1030	0.0422	0.0972	0.0337	0.0895	0.0286
	SRMQ1	0.1841	0.0913	0.1246	0.0849	0.1033	0.0364	0.0951	0.0285	0.0900	0.0238
	TLBP	0.2085	0.2063	0.1244	0.0912	0.1086	0.0576	0.0973	0.0395	0.0877	0.0376

, which also begins on the subsequent page. In Figure 4 and Table 3, under different payloads, “Original” represents the undetectability of the original images, while “Proposed” represents the undetectability of images with secret information embedded after applying our image vaccine.

Our experimental analysis yielded compelling insights into the efficacy of our method across a spectrum of steganographic algorithms. The findings underscore the superiority of our approach, which consistently surpasses alternative steganographic analysis techniques at all payload levels. While the enhancements at lower payloads are subtle, the benefits of employing the CLAHE algorithm become increasingly evident as the payload escalates, yielding heightened accuracy and robustness.

This pattern is especially noticeable in the defense against WOW, S-UNIWARD, and MiPOD algorithms, where our method demonstrates a more stable and pronounced performance advantage at medium-to-high payloads. The effectiveness of our approach is further highlighted in its handling of the intricate UT-GAN steganographic method, where it achieves a notable detection edge at lower payloads, exemplified by a reduction in the detection error rate from 0.2598 to 0.1379 at 0.1 bpp. This reduction signifies a heightened sensitivity to steganographic information crafted by adversarial networks.

This improvement is because our method changes the entropy of the images, resulting in a more regular histogram pattern. Such changes make steganographic operations easier to detect, as the increased regularity often reveals hidden information patterns, thereby enhancing the accuracy of steganalysis. In conclusion, our method exhibits superior performance across various steganography algorithms and payload rates, confirming its effectiveness and broad applicability against steganographic attacks.

It is important to acknowledge that our method does not significantly improve the detection of conventional steganographic algorithms at a low embedding rate of 0.1 bpp. This is likely because the subtle impact on local image features makes the effect of CLAHE less pronounced. However, even modest improvements are crucial for practical image security applications. This highlights the reliability of our “universal image vaccine” across various payload levels, confirming its broad utility in digital image security.

To validate the performance of our proposed method across different datasets, we conducted comparisons using the BOSSbase ver. 1.01 dataset with 10,000 images, focusing on four different steganography methods at a payload rate of 0.4 bpp, as illustrated in Figure 5. The experimental results indicate that our method significantly reduces the values of $P_E$ and greatly enhances the detection performance of steganalysis algorithms for all tested steganography methods. Particularly, in the case of the Yedroudj-Net algorithm, our proposed method achieves a 100% detection rate for all four types of steganography. These results confirm the robustness and effectiveness of our method across different datasets. The specific values for Figure 5 are detailed in

Table 4. Impact of proposed method on $P_E$ with BOSSbase v1.01.

Steganography (0.4 bpp)	Steganalysis	Original	Proposed
WOW	SPAM	0.3700	0.1420
	SRM	0.2565	0.1463
	SRMQ1	0.2688	0.1265
	TLBP	0.2834	0.1978
	Yedroudj-Net	0.1940	0.0
SUNIWARD	SPAM	0.3370	0.1106
	SRM	0.2465	0.1097
	SRMQ1	0.2524	0.1121
	TLBP	0.2647	0.1516
	Yedroudj-Net	0.1775	0.0
MiPOD	SPAM	0.3831	0.1167
	SRM	0.2934	0.1201
	SRMQ1	0.3032	0.1027
	TLBP	0.3075	0.1545
	Yedroudj-Net	0.2570	0.0
UT-GAN	SPAM	0.1993	0.0129
	SRM	0.1118	0.0151
	SRMQ1	0.1155	0.0124
	TLBP	0.1235	0.0372
	Yedroudj-Net	0.0910	0.0

.

To further investigate the differences between our proposed method and existing deep learning-based steganalysis techniques, we conducted a comparative analysis with Yedroudj-Net under varying payloads. We meticulously controlled experimental parameters, such as image dimensions, formats, and steganographic payloads, to minimize their impact on the results. As shown in Figure 6a, within the BOSSbase ver 1.01 dataset containing 10,000 images, the optimization effect on Yedroudj-Net varies with different payloads. Higher payloads, such as 0.4 and 0.5, result in significant improvements in detection accuracy due to the “Image Vaccine”. We used the CLAHE algorithm to enhance local contrast and regularize the histogram. This process may aid steganalysis algorithms in detecting anomalous patterns, thereby improving detection precision. As a deep learning model, Yedroudj-Net is potentially more sensitive to subtle image changes, allowing it to better identify steganographic content in images processed with the “Image Vaccine”.

Additionally, we examined the impact of different numbers of digital images on our method. Figure 6b presents experiments based on varying sample sizes within the BOWS2 dataset, evaluating our method’s effectiveness against the S-UNIWARD steganography and comparing it with the SRMQ1 steganalysis method. Clearly, whether for 1000 or 9000 images, our method consistently maintained a significantly lower error rate across all tested data volumes, indicating its practical applicability and scalability in handling different data scales.

4. Conclusions

In this study, we explored strategies for the protective processing of digital images to counter steganographic attacks and introduced a “universal image vaccine” scheme. By utilizing the CLAHE algorithm, we enhanced the detection capabilities of existing steganographic analysis systems and improved image visual quality. Our experimental results demonstrated that our method effectively countered both traditional and machine learning-based steganographic attacks, particularly at medium-to-high payloads, significantly enhancing detection capabilities. Specifically, for payloads of 0.1–0.5 bpp, the detection error rate was reduced by an average of 13.43%, 24.70%, 33.79%, 44.50%, and 51.06% for traditional algorithms. Notably, when countering the MiPOD steganographic algorithm at a payload of 0.4 bpp, the Yedroudj-Net algorithm showed a dramatic reduction in detection error rate, decreasing from 0.2570 to 0.0000, achieving 100% detection accuracy and highlighting the robust anti-steganographic performance of our approach. This finding suggested potential applications, such as embedding the processing algorithm into chips, to bolster the security of digital images.

While our approach is intended for broad applicability rather than being customized for specific research contexts, we recognize that its optimization at low embedding rates, particularly at 0.1 bpp, does not match the effectiveness observed at higher bit rates. Moreover, with the advancement of steganography and increasingly complex embedding techniques, future robust algorithms may present new challenges. However, we note that there is a growing body of research dedicated to steganalysis in the low-embedding-rate domain [37,38]. The integration of such algorithms could lead to a more holistic enhancement in performance.

Subsequent research will investigate the enhancement of anti-steganographic efficacy at low payloads and the development of specialized processing methodologies applicable to a range of image modalities, including medical imaging. This endeavor will contribute to the development of an advanced defense system that integrates a range of image vaccine technologies. Furthermore, we aim to assess the adaptability and efficiency of our “universal image vaccine” across diverse applications and its compatibility with existing steganographic analysis tools, with the goal of offering comprehensive solutions to the challenges posed by steganography.